Beyond BMI: Unsupervised Machine Learning Clustering of Bioelectrical Impedance Phenotypes Is Associated With...

To identify data-driven body composition phenotypes in patients with moderate-to-extreme anorexia nervosa (AN) using machine learning (ML) clustering of bioelectrical impedance analysis (BIA) parameters, and to test whether these phenotypes have specific psychopathological and childhood trauma profiles while correlating with early weight-recovery trajectories.

In a prospective observational cohort, 100 females (18-40 years) with AN and body mass index (BMI) < 17 kg/m² were enrolled. Baseline BIA-derived measures (phase angle, body cell mass, fat mass, fat-free mass, total body water, and extracellular water) were clustered using Unsupervised Random Forest algorithms and consensus-based selection of the optimal number of clusters. Between-cluster differences were tested, and longitudinal BMI trajectories were modeled with generalized additive mixed models across weekly follow-ups up to 3 months.

Three clusters emerged (n = 23, 53, 24) with separation driven primarily by hydration and lean-mass indices (total body water, fat-free mass, body cell mass), yielding phenotypes consistent with "Preserved Body Cell Mass", "Severe Depletion", and "Fluid Redistribution" (elevated extracellular water with low body cell mass and phase angle), despite overlapping BMI. Clusters 1 and 2 showed higher childhood trauma exposure and specific trait differences. At 3 months, Cluster 3 showed a more pronounced BMI increase over time than Clusters 1 and 2.

BIA-based ML phenotyping may complement current BMI-centric severity staging by capturing distinct morpho-functional patterns associated with differential early weight-recovery trajectories. BIA provides a noninvasive low-cost assessment, supporting more personalized nutritional and psychotherapeutic planning and informing future precision-staging research in AN.